The Thermodynamic AI Chip · Thomas Ahle

2026-06-28 · Source: Machine Learning Street Talk · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

Normal Computing, led by Thomas Ahle, is developing "thermodynamic" AI chips, exemplified by their CN 101 silicon release, which utilize inherent noise for probabilistic computation. This approach aims to make chips behave as stochastic differential equations, potentially accelerating Bayesian inference and diffusion models. Ahle's team also uses AI agents to automate chip design, generating over 500,000 lines of Verilog code for a simulator in 43 days, circumventing commercial software costs of \$10,000 per CPU kernel. However, this raises concerns about the "spaghetti monster" problem of unstructured AI-generated code and the difficulty of formal verification, as models often achieve 70-80% test pass rates but rarely 100% correctness. The discussion also touches on autoformalization for chip design and the broader societal impact of AI on human understanding and collaboration.

Key takeaway

For AI Hardware Engineers evaluating advanced design methodologies, recognize that while AI agents can drastically reduce chip design costs and accelerate Verilog code generation, they introduce significant verification challenges due to code complexity. Prioritize robust formal verification tools and strategies to maintain design integrity. Consider exploring thermodynamic computing for specialized probabilistic workloads, but be mindful of the need for new algorithms to fully exploit such novel architectures.

Key insights

AI-driven hardware design offers efficiency gains but demands robust verification and strategies to preserve human understanding.

Principles

• Noise can be harnessed as computation in specialized hardware.
• Formal verification is paramount for complex chip designs.
• AI-generated code often lacks human legibility and structure.

Method

AI agents collaborate to generate Verilog code for chip design, followed by formal verification. Thermodynamic computing infuses noise into programmable arrays, allowing chips to solve stochastic differential equations by settling into answers.

In practice

• Design custom circuits for specific algorithms with AI.
• Optimize and verify chip designs using AI tools.
• Apply thermodynamic chips to probabilistic workloads.

Topics

• AI Hardware Design
• Formal Verification
• Thermodynamic Computing
• AI Code Generation
• Verilog Simulation
• Probabilistic Workloads

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Hardware Engineer, AI Scientist, AI Engineer

Related on AIssential

• Thermonat: A machine learning tool for semiconductor modeling — High-accuracy, high-speed thermal simulation of individual transistors is crucial for advanced circuit design.
• AI Agent Designs a RISC-V CPU Core From Scratch — Agentic AI systems can autonomously design complete CPU cores, mirroring human engineering workflows.
• The Sequence Opinion #819: How AI Chips are Made? — Hardware, not just algorithms, drives 60% of deep learning's current success.
• A new version of OpenAI’s Codex is powered by a new dedicated chip — Dedicated AI hardware accelerates lightweight models for real-time developer collaboration and rapid iteration.
• Cognichip wants AI to design the chips that power AI, and just raised $60M to try — AI can significantly accelerate and de-risk complex chip design by automating and optimizing engineering workflows.
• IBM, Synopsys Move Toward 1.4-nm Node With Heat-Modeling Tech — A new ML-powered thermal modeling tool accelerates nanoscale chip design, enabling 1.4-nm nodes and beyond.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning Street Talk.